The Blind Spot
Only 15% of generative AI deployments currently have observability in place. The remaining 85% are running AI agents in production without real visibility into whether outputs are accurate, safe, or costing what they should. Gartner predicts this will rise to 50% by 2028, but that still leaves half of all deployments flying blind two years from now. [1]
The gap is not for lack of tooling. Gartner published its first Market Guide for AI Evaluation and Observability Platforms in February 2026, listing over a dozen vendors. Braintrust closed an $80 million Series B at an $800 million valuation. Langfuse was acquired by ClickHouse. The market is growing at over 36% annually. [2]
The problem is that most engineering teams do not yet understand what agent observability is, why traditional monitoring cannot substitute for it, or where to start.
Why Traditional Monitoring Fails
Traditional application performance monitoring assumes deterministic request-response cycles. A request enters, code executes, a response exits. If latency spikes or the status code is not 200, something is wrong. AI agents break every one of these assumptions.
The same prompt can produce different tool calls on different runs. The execution path branches based on model output, not code logic. A 200 response can wrap a confidently wrong answer. An agent that selects the wrong API tool during a multi-step workflow reports a successful HTTP request while hallucinated responses reach customers. [3]
Agent observability does not replace traditional monitoring — it adds the layer that explains the chain of decisions that produced the outcome. Where normal monitoring observes outputs, agent observability explains the reasoning path that generated them.
Traces, Spans, and Evals
The unit of agent observability is the span. A span captures a single operation: one LLM call, one tool invocation, one retrieval query, one agent reasoning step. Spans carry inputs, outputs, latency, token counts, and metadata. Spans nest into traces, which represent the full decision path of a single agent run.
OpenTelemetry's GenAI Semantic Conventions define a standard span model for AI agents. An invoke_agent span contains chat spans (each model call), which contain execute_tool spans (each tool invocation). As of mid-2026, every GenAI attribute and span in the specification carries the "Development" stability badge — none has reached Stable — but the convention is already supported by Datadog, Honeycomb, New Relic, and major agent frameworks like LangChain and CrewAI.